1. Whirlwind Tour of Hadoop Edward Capriolo Rev 2

2. Whirlwind tour of Hadoop Inspired by Google's GFS Clusters from 1-10000 systems Batch Processing High Throughput Partition-able problems Fault Tolerance in storage Fault Tolerance in processing

3. HDFS Distributed File System (HDFS) Centralized “INODE” store NameNode DataNodes provide storage 1-10000 Replication set Per File Block Size normally large 128 MB Optimized for throughput

4. HDFS

5. Map Reduce Jobs are typically long running Move processing not data Algorithm uses map() and optionally reduce() Source Input is split Splits processed simultaneously Failed Tasks are retried

6. MapReduce

7. Things you do not have Locking Blocking Mutexes wait/notify POSIX file semantics

8. How to: Kick rear Slave Nodes Lots of Disk (no RAID) – Storage – More Spindles Lots of RAM 8+ GB Lots of CPU/Cores

9. Problem: Large Scale Log Processing Challenges Large volumes of Data Store it indefinitely Process large data sets on multiple systems Data and Indexes for short periods (one week) do NOT fit in main memory

10. Log Format

11. Getting Files into HDFS

12. Hadoop Shell Hadoop has a shell ls, put, get, cat..

13. Sample Program Group and Count Source data looks like jan 10 2009:.........:200:/index.htm jan 10 2009:.........:200:/index.htm jan 10 2009:.........:200:/igloo.htm jan 10 2009:.........:200:/ed.htm

14. In case your a Math 'type' (input) → map -> -> combine -> -> reduce -> (output) Map(k1,v1) -> list(k2,v2) Reduce(k2, list (v2)) -> list(v3)

15. Calculate Splits Input does not have to be a file to be “splittable”

16. Splits → Mappers Mappers run the map() process on each split

17. Mapper runs map function

18. Shuffle sort, NOT the hottest new dance move Data Shuffled to reducer on key.hash() Data sorted by key per reducer.

19. Reduce ”index.htm”, “ed,htm”,

20. Reduce Phase

21. Scalable Tune mappers and reducers per job More Data...get...More Systems Lose DataNode files replicate elsewhere Lose TaskTracker running tasks will restart Lose NameNode, Jobtracker- Call bigsaad

22. Hive Translates queries into map/reduce jobs SQL like language atop Hadoop like MySQL CSV engine, on Hadoop, on roids Much easier and reusable then coding your own joins, max(), where CLI, Web Interface

23. Hive has no “Hive Format” Works with most input Configurable delimiters Types like string,long also map, struct, array Configurable SerDe

24. Create & Load hive>create table web_data ( sdate STRING, stime STRING, envvar STRING, remotelogname STRING,servername STRING, localip STRING, literaldash STRING, method STRING, url STRING, querystring STRING, status STRING, litteralzero STRING,bytessent INT,header STRING, timetoserver INT, useragent STRING,cookie STRING, referer STRING); LOAD DATA INPATH '/user/root/phpbb/phpbb.extended.log.2009- 12-19-19' INTO TABLE web_data

25. Query SELECT url,count(1) FROM web_data GROUP BY url Hive translates query into map reduce program(s)

29. The End... Questions???